diff --git a/Dockerfile b/Dockerfile
new file mode 100644
index 0000000..b3ee9f5
--- /dev/null
+++ b/Dockerfile
@@ -0,0 +1,71 @@
+# syntax=docker/dockerfile:1.7
+#
+# Targets:
+#   runtime — vLLM only, model from HF cache at runtime (default)
+#   serve   — runtime + model baked in
+#
+# Runtime (default):
+#   docker build -t moss-td-vllm:dev .
+#
+# Serve (reuse local HF cache, no re-download):
+#   docker build --target serve \
+#     --build-context models=$HOME/.cache/huggingface \
+#     --build-arg MODEL_REVISION=e5118b411bf5a77d7a90c4941066bec93c967312 \
+#     -t moss-td-serve:1.0.0 .
+
+ARG CUDA_IMAGE=nvidia/cuda:13.3.0-cudnn-devel-ubuntu24.04
+FROM ${CUDA_IMAGE} AS runtime
+
+ENV DEBIAN_FRONTEND=noninteractive \
+    PYTHONUNBUFFERED=1 \
+    PIP_DISABLE_PIP_VERSION_CHECK=1 \
+    PIP_NO_CACHE_DIR=1 \
+    UV_HTTP_TIMEOUT=600
+
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    ca-certificates \
+    curl \
+    ffmpeg \
+    git \
+    libsndfile1 \
+    python3.12 \
+    python3.12-dev \
+    python3-pip \
+    && rm -rf /var/lib/apt/lists/*
+
+RUN curl -LsSf https://astral.sh/uv/install.sh | sh
+
+ENV PATH="/root/.local/bin:${PATH}"
+# README: CUDA 13 → cu130 pinned nightly
+ARG VLLM_CUDA=cu130
+ARG VLLM_WHEEL_COMMIT=68b4a1d582818e67adc903bf1b8fc5a5447da2fa
+
+# Install the CUDA PyTorch wheels explicitly so uv does not resolve to torch+xpu.
+# Pin to the 2.11 release family to match the CUDA 13 wheels used by the project docs.
+RUN uv pip install --system --break-system-packages -U \
+    torch==2.11.0 \
+    torchvision==0.26.0 \
+    torchaudio==2.11.0 \
+    --python /usr/bin/python3.12 \
+    --index-url "https://download.pytorch.org/whl/${VLLM_CUDA}"
+
+RUN uv pip install --system --break-system-packages -U "vllm[audio]" \
+    --python /usr/bin/python3.12 \
+    --extra-index-url "https://wheels.vllm.ai/${VLLM_WHEEL_COMMIT}/${VLLM_CUDA}"
+
+WORKDIR /workspace
+EXPOSE 8000
+CMD ["vllm", "serve", "OpenMOSS-Team/MOSS-Transcribe-Diarize", "--trust-remote-code", "--host", "0.0.0.0", "--port", "8000"]
+
+# Model baked in; reuses runtime layer cache when only MODEL_REVISION changes.
+FROM runtime AS serve
+
+ARG MODEL_REPO=models--OpenMOSS-Team--MOSS-Transcribe-Diarize
+ARG MODEL_REVISION=e5118b411bf5a77d7a90c4941066bec93c967312
+
+# HF snapshots are symlinks into hub/blobs; cp -aL dereferences them into real files.
+RUN --mount=type=bind,from=models,source=/,target=/hf,ro \
+    mkdir -p /models/MOSS-Transcribe-Diarize && \
+    cp -aL "/hf/hub/${MODEL_REPO}/snapshots/${MODEL_REVISION}/." /models/MOSS-Transcribe-Diarize/
+
+CMD ["vllm", "serve", "/models/MOSS-Transcribe-Diarize", "--served-model-name", "OpenMOSS-Team/MOSS-Transcribe-Diarize", "--trust-remote-code", "--host", "0.0.0.0", "--port", "8000"]
diff --git a/docker-compose.yml b/docker-compose.yml
new file mode 100644
index 0000000..fa50363
--- /dev/null
+++ b/docker-compose.yml
@@ -0,0 +1,58 @@
+# vLLM serving for MOSS-Transcribe-Diarize
+#
+# Build & run:
+#   docker compose up --build
+#
+# GPU (runtime env vars, not build args):
+#   GPUS=all docker compose up --build
+#   GPUS=1 docker compose up --build
+#   NVIDIA_VISIBLE_DEVICES=0,1 docker compose up --build
+#
+# Image tag:
+#   IMAGE_TAG=1.0.0 docker compose up --build
+#
+# Serve image with model baked in:
+#   docker compose --profile serve up --build
+
+services:
+  vllm:
+    build:
+      context: .
+      dockerfile: Dockerfile
+      target: runtime
+    image: ${IMAGE_NAME:-moss-td-vllm}:${IMAGE_TAG:-dev}
+    user: root
+    volumes:
+      - ${PWD}:/workspace
+      - ${HF_HOME:-${HOME}/.cache/huggingface}:/root/.cache/huggingface
+    working_dir: /workspace
+    ports:
+      - "8000:8000"
+    shm_size: "16gb"
+    environment:
+      NVIDIA_VISIBLE_DEVICES: ${NVIDIA_VISIBLE_DEVICES:-all}
+      NVIDIA_DRIVER_CAPABILITIES: ${NVIDIA_DRIVER_CAPABILITIES:-compute,utility}
+    gpus: ${GPUS:-all}
+    stdin_open: true
+    tty: true
+
+  serve:
+    profiles: [serve]
+    build:
+      context: .
+      dockerfile: Dockerfile
+      target: serve
+      additional_contexts:
+        models: ${HF_HOME:-${HOME}/.cache/huggingface}
+      args:
+        MODEL_REVISION: ${MODEL_REVISION:-e5118b411bf5a77d7a90c4941066bec93c967312}
+    image: ${SERVE_IMAGE:-moss-td-serve}:${IMAGE_TAG:-latest}
+    user: root
+    ports:
+      - "8000:8000"
+    shm_size: "16gb"
+    environment:
+      NVIDIA_VISIBLE_DEVICES: ${NVIDIA_VISIBLE_DEVICES:-all}
+      NVIDIA_DRIVER_CAPABILITIES: ${NVIDIA_DRIVER_CAPABILITIES:-compute,utility}
+    gpus: ${GPUS:-all}
+    restart: unless-stopped